The Resolution Illusion: How AI Super-Resolution is Rewriting the Rules of the Aerial Imagery Market
For decades, the manned aerial imagery market operated on a straightforward premise: acquiring imagery sharp enough to identify roof damage, count vehicles, or measure field boundaries to centimeter accuracy required aircraft, pilots, sensors, and substantial capital. That assumption is now under pressure from an unexpected source—satellite imagery combined with generative AI.
On May 5th, 2026, Planet launched Planet SuperRes, an AI product that enhances PlanetScope's native 3-meter resolution to an effective 2 meters with near-daily global coverage. Around the same time, the European Space Agency's freely available SEN2SR framework demonstrated Sentinel-2 enhancement to 2.5 meters, and independent researcher Yosef Akhtman's S2DR3 model claimed 1-meter reconstruction from 10-meter Sentinel-2 input. The question for aerial imagery companies, their investors, and their premium-price customers: does this actually change the market?
What Super-Resolution Actually Is (And Isn't)
Super-resolution uses deep neural networks—almost universally Generative Adversarial Networks (GANs)—to reconstruct spatial detail beyond what the sensor physically captured. A Generator network attempts to produce synthetic high-resolution imagery from low-resolution input; a Discriminator network tries to tell that synthetic output apart from genuine high-resolution captures. Trained against each other, the Generator gets pushed toward output that's perceptually indistinguishable from the real thing.
Planet's SuperRes runs an Enhanced Super-Resolution GAN (ESRGAN) trained on roughly 120,000 matched image pairs from its SkySat and PlanetScope constellations, and ships with a pixel-level confidence layer so analysts can judge how trustworthy a given reconstruction is. This matters strategically: a GAN's quality ceiling depends almost entirely on the size and quality of its paired training data, not on architecture. The open-source GAN code is freely available; a 120,000-image archive of matched high-and-low-resolution pairs is not. That's precisely why decades of sub-10cm EagleView, Nearmap, and Vexcel archive imagery just became extraordinarily valuable training data in a GAN-dominated world.
Super-resolution is a resolution amplifier with bounded confidence—not a substitute for capture.
Where It Works, and Where It Breaks Down
Regional land cover monitoring, agricultural field boundaries, vegetation assessment, and broad change detection all benefit meaningfully from satellite super-resolution. One-meter effective resolution with near-daily revisit beats a manned survey on a six-month refresh cycle for these use cases.
Insurance underwriting and property condition assessment depend on radiometrically calibrated, 5–7.5cm GSD imagery precise enough to identify roof material, detect active water retention, or measure slope angle. Super-resolution can't reconstruct spectral and spatial detail that was never captured at the source—it's mathematical inference, not fabrication.
Legal admissibility adds another layer. Where imagery serves as evidence—claims, disputes, infrastructure assessments—manned capture's data provenance and geometric certification still substantially outperforms AI-reconstructed satellite imagery in court and regulatory contexts. That's true today. It won't necessarily stay true.
The Threat Model: Who's Actually at Risk
The displacement risk is not evenly distributed across the industry, and it does not fall where most people assume.
Companies like NV5 and Sanborn compete project-by-project for state DOT, county assessor, and utility contracts—frequently at 15–30cm resolution for topographic mapping and base-layer refresh. That's exactly the resolution band where satellite-plus-super-resolution now offers a credible, dramatically cheaper alternative, and with no subscription stickiness, every lost RFP simply disappears.
Planet's near-daily global coverage at 2-meter effective resolution now beats scheduled manned flights on speed as well as cost for most non-urgent monitoring work.
ESA's SEN2SR is free under a Creative Commons Zero license. Any technically capable utility, agency, or nonprofit can point it at cost-free Sentinel-2 data and generate 2.5-meter imagery without paying a commercial vendor—a direct challenge to mid-tier suppliers selling products sophisticated customers can now replicate at near-zero marginal cost.
Notably absent from that list: Nearmap and Vexcel. Their sub-10cm subscription businesses, embedded deep in insurance underwriting workflows, sit on the other side of a resolution and certification gap that satellite super-resolution isn't close to closing.
Sixty Carlton's View: The durable moat in manned aerial imagery has never been resolution alone—it's resolution delivered with certainty, certification, and an auditable data-provenance chain that automated enterprise decisioning demands. Planet's confidence layer is smart product design, but it also implicitly draws the boundary of where satellite-plus-AI is welcome: human-in-the-loop analysis, yes; closed-loop automated underwriting, not yet. That boundary is exactly where the next twelve months of competitive positioning will be fought.
Three Ways to Play the Interval
Super-resolution's resolution ceiling will keep rising as training data and architectures mature—image segmentation and object detection both faced similar "not good enough yet" skepticism a decade ago. The real strategic question isn't whether the gap closes, but what a geospatial company does in the interval before it does.
Oblique capture, sub-10cm multispectral imaging, true-ortho generation from elevation models, and interior/exterior property workflows aren't reachable by satellite super-resolution. Depth here is a structurally sound bet.
Planet's SuperRes is only as good as its SkySat training corpus. Decades of sub-10cm archive imagery held by EagleView, Nearmap, and Vexcel is exactly the paired ground-truth data that satellite super-resolution vendors need and don't have. Licensing it converts a potential competitor into a customer.
Insurance and mortgage industries will eventually need guidance on AI-enhanced imagery in underwriting. Whoever helps write what "certified imagery" means in an AI-augmented world gets first-mover advantage when those standards crystallize.
Is your geospatial data strategy still competing on resolution alone, or are you building the provenance and training-data moat that survives the next model generation? Let's talk.